Low-Cost and Fast Epiretinal Membrane Detection and Quantification Based on SD-OCT

Seungju Baek, Insup Lee, Kuk Jin Jang, Yongseop Han, Jin Hyun KimIEEE Access 2025 · DOI

TL;DR

This paper introduces a novel, cost-effective pipeline for detecting and quantifying epiretinal membranes (ERMs) using standard Spectral-Domain Optical Coherence Tomography (SD-OCT) B-scans. By leveraging deep learning (YOLOv11x) and a new "Epiretinal Projection Image" (EPI) technique, the method achieves high accuracy and provides intuitive visualization, bridging the gap between affordable SD-OCT and more advanced but expensive Swept-Source OCT (SS-OCT) systems.

What Problem Does This Paper Address?

Epiretinal membranes (ERMs) are a common cause of vision impairment, and their diagnosis and management are crucial. Optical Coherence Tomography (OCT) is the primary imaging tool for this. While Spectral-Domain OCT (SD-OCT) is widely available and affordable, its B-scan images provide only cross-sectional views, making it challenging to assess the full extent and morphology of ERMs. Swept-Source OCT (SS-OCT) offers superior en face visualization, which is more intuitive for ERM assessment, but its high cost and complexity limit its clinical adoption. This creates a need for a solution that combines the accessibility of SD-OCT with the diagnostic clarity of SS-OCT for ERM detection and quantification.

Methods

The proposed pipeline consists of several key steps:

  • ERM Detection: A YOLOv11x object detection model is trained to identify ERM regions within individual SD-OCT B-scan images. The authors experimented with YOLOv5, YOLOv8, and YOLOv11, finding YOLOv11x to be the most effective.
  • Epiretinal Projection Image (EPI) Generation: Detected ERM locations from multiple B-scans (both horizontal and vertical) are normalized and projected onto a single en face plane. This creates the EPI, which provides a comprehensive, high-resolution visualization of the ERM's spatial distribution and morphology. This is a novel contribution.
  • ERM Area Quantification: The total ERM area is calculated by counting pixels within the EPI.
  • Association Quantification: A novel method quantifies the relationship between the ERM area (from the EPI) and retinal thickness maps. This involves discretizing retinal thickness into nine classes and calculating an "Association Score" (ASSscore) based on the overlap between ERM and thickened retinal areas, weighted by the clinical significance of thickness. This allows for a more objective assessment of ERM's impact on retinal thickness.

Datasets

The study utilized a dataset of 2,850 SD-OCT B-scan images. These images were annotated by ophthalmologists. For training and evaluation, two sets of training data were used: 1,300 images and 2,200 images. A common evaluation set of 650 images was used across all experiments. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Bridging the SD-OCT/SS-OCT Gap: The EPI technique effectively simulates the intuitive en face visualization of SS-OCT using readily available SD-OCT data, offering a cost-effective alternative.
  • YOLOv11x Superiority: Among the tested YOLO versions, YOLOv11x demonstrated the highest detection performance, indicating its suitability for complex medical image analysis.
  • Data Size Matters: Increasing the training dataset size from 1,300 to 2,200 images consistently improved the performance of all YOLO models, highlighting the importance of data sufficiency.
  • ERM-Retinal Thickness Correlation: The novel association quantification method provides a data-driven way to understand how ERM impacts retinal thickness, which is crucial for treatment planning and can help differentiate ERM-induced changes from other pathologies.
  • Clinical Validation: The model's outputs were evaluated by board-certified retina specialists, showing excellent agreement and high clinical acceptability, suggesting its potential for real-world use.

Key Results

  • ERM Detection Accuracy: The YOLOv11x model achieved high detection accuracy, with mAP@50 of 0.882 and mAP@50:95 of 0.556.
  • Clinical Agreement: Inter-expert agreement for the model's outputs was excellent, with an ICC of 0.94 and Cohen's kappa (κ) of 0.89.
  • Clinical Acceptability: The average Likert score for clinical accuracy and acceptability was 4.48 out of 5, with an Acceptability Rate (AR) of 85.0%, indicating strong expert confidence.
  • Association Score: The correlation score between ERM presence and retinal thickening was found to be 0.771, demonstrating a strong relationship.

Limitations and Future Work

  • Limited Expert Evaluation: The clinical validation involved only two board-certified retina specialists, which may not fully represent the variability of interpretations across a broader population. Future work should involve a larger and more diverse cohort of specialists.
  • Ordinal Data Constraints: The use of a 5-point Likert scale provides ordinal data, which has statistical limitations for parametric analyses.
  • ERM Boundary Delineation: Accurately delineating ERM boundaries in OCT images remains inherently difficult due to inter-patient variability. The current study focused on model reliability and clinical validity through expert consensus rather than establishing an absolute ground truth.
  • Sample Size: The evaluation of 30 ERM cases, while providing a preliminary basis, is a limited sample size. Future studies with larger, multi-center cohorts are recommended.
  • Segmentation Refinement: The authors suggest future integration with segmentation-based methods (e.g., Mask R-CNN, U-Net) to improve morphological fidelity and reduce potential underestimation of ERM area and length, especially for complex membrane shapes.
  • Direct SS-OCT Comparison: While the method aims to replicate SS-OCT advantages, a direct, objective comparison with SS-OCT-based quantification is still needed. Future work will explore this.

Summary updated

BibTeX
@article{baek2025low,
  author = {Seungju Baek and Insup Lee and Kuk Jin Jang and Yongseop Han and Jin Hyun Kim},
  title = {{Low-Cost and Fast Epiretinal Membrane Detection and Quantification Based on SD-OCT}},
  journal = {IEEE Access},
  year = {2025},
  doi = {10.1109/access.2025.3629332},
}
Figure